When the regime shifts but the signal does not
A practical framework for separating a durable factor from the market environment that made it look durable.
Most systematic strategies begin with a stable sentence: when this happens, that tends to follow. The sentence is then translated into a signal, tested across a history, and eventually entrusted with capital. The difficulty is that markets do not share our preference for stable sentences.
A signal can remain measurable while the mechanism that supports it changes. Equally, a sound mechanism can pass through an environment in which its realised returns look indistinguishable from failure. Treating these as the same problem leads either to stubbornness or overfitting.
Separate the signal from its weather
The first task is to distinguish explanatory variables from conditioning variables. The former describe why the premium should exist. The latter describe the weather in which we are trying to harvest it: liquidity, dispersion, volatility, crowding and the price of balance-sheet capacity.
This distinction matters because the two classes deserve different responses. Evidence against the economic mechanism should reduce our belief in the signal. A hostile conditioning environment should affect the amount of risk we are prepared to take.
“A regime model is most useful when it makes the portfolio less certain, not when it pretends to make the future certain.”
A three-part diagnostic
I use three tests. First, ask whether the cross-sectional ordering still appears before costs. Second, measure whether the opportunity set has compressed or merely become noisier. Third, look for evidence that implementation—not prediction—has become the binding constraint.
- Mechanism: is the behavioural or structural reason for the return still observable?
- Opportunity: is there enough dispersion for the signal to express itself?
- Translation: are costs, liquidity or crowding absorbing the forecast?
None of these produces a single regime label. That is intentional. The output is a set of conditional beliefs, which is a more honest input to portfolio construction than a brightly coloured state machine.
What changes in the portfolio
When mechanism evidence weakens, the forecast itself should decay. When the opportunity set contracts, risk should fall while the research prior remains intact. And when translation deteriorates, the remedy belongs in implementation: slower turnover, broader baskets, or a higher hurdle for marginal trades.
The distinction is not semantic. It prevents a period of poor returns from becoming permission to change every moving part at once. Good research keeps an audit trail between observation, diagnosis and action.